IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend...

14
E-MiLi: Energy-Minimizing Idle Listening in Wireless Networks Xinyu Zhang, Student Member, IEEE, and Kang G. Shin, Fellow, IEEE Abstract—WiFi interface is known to be a primary energy consumer in mobile devices, and idle listening (IL) is the dominant source of energy consumption in WiFi. Most existing protocols, such as the 802.11 power-saving mode (PSM), attempt to reduce the time spent in IL by sleep scheduling. However, through an extensive analysis of real-world traffic, we found more than 60 percent of energy is consumed in IL, even with PSM enabled. To remedy this problem, we propose Energy-Minimizing idle Listening (E-MiLi) that reduces the power consumption in IL, given that the time spent in IL has already been optimized by sleep scheduling. Observing that radio power consumption decreases proportionally to its clock rate, E-MiLi adaptively downclocks the radio during IL, and reverts to full clock rate when an incoming packet is detected or a packet has to be transmitted. E-MiLi incorporates sampling rate invariant detection, ensuring accurate packet detection and address filtering even when the receiver’s sampling clock rate is much lower than the signal bandwidth. Further, it employs an opportunistic downclocking mechanism to optimize the efficiency of switching clock rate, based on a simple interface to existing MAC-layer scheduling protocols. We have implemented E-MiLi on the USRP software radio platform. Our experimental evaluation shows that E-MiLi can detect packets with close to 100 percent accuracy even with downclocking by a factor of 16. When integrated with 802.11, E-MiLi can reduce energy consumption by around 44 percent for 92 percent of users in real-world wireless networks. Index Terms—Energy efficiency, CSMA wireless networks, idle listening, packet detection, adapting clock rate, dynamic frequency scaling. Ç 1 INTRODUCTION C ONTINUING advances of physical-layer technologies have enabled WiFi to support high data rates at low cost and hence become widely deployed in networking infrastruc- tures and mobile devices, such as laptops, smartphones, and tablet PCs. Despite its high performance and inexpensive availability, the energy efficiency of WiFi remains a challen- ging problem. For instance, WiFi accounts for more than 10 percent of the energy consumption in current laptops [1]. It may also raise a smartphone’s power consumption 14 times even without packet transmissions [2]. WiFi’s energy inefficiency comes from its intrinsic CSMA mechanism—the radio must perform idle listening (IL) continuously, in order to detect unpredictably arriving packets or assess a clear channel. The energy consumption of IL, unfortunately, is comparable to that of active transmission/reception [2], [3]. Even worse, WiFi clients tend to spend a large fraction of time in IL, due to MAC- level contention and network-level delay [4]. Therefore, minimizing the IL’s energy consumption is crucial to WiFi’s energy efficiency. A natural way to reduce the IL’s energy cost is sleep scheduling. In WiFi’s power-saving mode (PSM) and its variants [1], [4], [5], [6], clients can sleep adaptively, and wake up only when they intend to transmit, or expect to receive packets. The AP buffers downlink packets and transmits only after the client wakes up. PSM essentially shapes the traffic by aggregating downlink packets, thereby reducing the receiver’s wait time caused by the network- level latency. However, it cannot reduce the IL time associated with carrier sensing and contention. Through an extensive trace-based analysis of real WiFi networks (Section 3), we have found that IL still dominates the clients’ energy consumption even with PSM enabled: it accounts for more than 80 percent of energy consumption for clients in a busy network and 60 percent in a relatively idle network. Since the IL time cannot be reduced any further due to WiFi’s CSMA, we exploit an additional dimension—redu- cing IL power consumption—in order to minimize its energy cost. Ideally, if the exact idle period is known, the radio could be powered off or put to sleep during IL, and wake up and process packets on demand. However, due to the distributed nature of CSMA, the idle time between packets varies widely and unpredictably. Underestimation of an idle interval will waste energy, while an over- estimation causes the radio to drop all incoming packets during the sleep. So, one may raise an important question: “is it possible to put the radio in a subconscious mode, where it consumes little power and can still respond to incoming packets promptly?” We answer this question by proposing Energy- Minimizing idle Listening (E-MiLi) that reduces the clock rate of the radio during its IL period. The power consumption of digital devices is known to be proportional to their voltage square and clock rate [7], [8]. Theoretically, by reducing clock rate alone, E-MiLi reduces the IL’s power consumption linearly. IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012 1441 . The authors are with the Real-Time Computing Laboratory, Department of Electrical Engineering and Computer Science, The University of Michigan, 2260 Hayward St., Ann Arbor, MI 48109-2121. E-mail: {xyzhang, kgshin}@eecs.umich.edu. Manuscript received 12 Nov. 2011; revised 6 Feb. 2012; accepted 19 Apr. 2012; published online 1 May 2012. For information on obtaining reprints of this article, please send e-mail to: [email protected] and reference IEEECS Log Number TMCSI-2011-11-0607. Digital Object Identifier no. 10.1109/TMC.2012.112. 1536-1233/12/$31.00 ß 2012 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

Transcript of IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend...

Page 1: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

E-MiLi: Energy-Minimizing IdleListening in Wireless Networks

Xinyu Zhang, Student Member, IEEE, and Kang G. Shin, Fellow, IEEE

Abstract—WiFi interface is known to be a primary energy consumer in mobile devices, and idle listening (IL) is the dominant source of

energy consumption in WiFi. Most existing protocols, such as the 802.11 power-saving mode (PSM), attempt to reduce the time spent

in IL by sleep scheduling. However, through an extensive analysis of real-world traffic, we found more than 60 percent of energy isconsumed in IL, even with PSM enabled. To remedy this problem, we propose Energy-Minimizing idle Listening (E-MiLi) that reducesthe power consumption in IL, given that the time spent in IL has already been optimized by sleep scheduling. Observing that radiopower consumption decreases proportionally to its clock rate, E-MiLi adaptively downclocks the radio during IL, and reverts to full

clock rate when an incoming packet is detected or a packet has to be transmitted. E-MiLi incorporates sampling rate invariantdetection, ensuring accurate packet detection and address filtering even when the receiver’s sampling clock rate is much lower than

the signal bandwidth. Further, it employs an opportunistic downclocking mechanism to optimize the efficiency of switching clock rate,based on a simple interface to existing MAC-layer scheduling protocols. We have implemented E-MiLi on the USRP software radio

platform. Our experimental evaluation shows that E-MiLi can detect packets with close to 100 percent accuracy even withdownclocking by a factor of 16. When integrated with 802.11, E-MiLi can reduce energy consumption by around 44 percent for

92 percent of users in real-world wireless networks.

Index Terms—Energy efficiency, CSMA wireless networks, idle listening, packet detection, adapting clock rate, dynamic frequency

scaling.

Ç

1 INTRODUCTION

CONTINUING advances of physical-layer technologies haveenabled WiFi to support high data rates at low cost and

hence become widely deployed in networking infrastruc-tures and mobile devices, such as laptops, smartphones, andtablet PCs. Despite its high performance and inexpensiveavailability, the energy efficiency of WiFi remains a challen-ging problem. For instance, WiFi accounts for more than10 percent of the energy consumption in current laptops [1].It may also raise a smartphone’s power consumption 14times even without packet transmissions [2].

WiFi’s energy inefficiency comes from its intrinsic CSMAmechanism—the radio must perform idle listening (IL)continuously, in order to detect unpredictably arrivingpackets or assess a clear channel. The energy consumptionof IL, unfortunately, is comparable to that of activetransmission/reception [2], [3]. Even worse, WiFi clientstend to spend a large fraction of time in IL, due to MAC-level contention and network-level delay [4]. Therefore,minimizing the IL’s energy consumption is crucial to WiFi’senergy efficiency.

A natural way to reduce the IL’s energy cost is sleepscheduling. In WiFi’s power-saving mode (PSM) and itsvariants [1], [4], [5], [6], clients can sleep adaptively, andwake up only when they intend to transmit, or expect to

receive packets. The AP buffers downlink packets andtransmits only after the client wakes up. PSM essentiallyshapes the traffic by aggregating downlink packets, therebyreducing the receiver’s wait time caused by the network-level latency. However, it cannot reduce the IL timeassociated with carrier sensing and contention. Throughan extensive trace-based analysis of real WiFi networks(Section 3), we have found that IL still dominates the clients’energy consumption even with PSM enabled: it accounts formore than 80 percent of energy consumption for clients in abusy network and 60 percent in a relatively idle network.

Since the IL time cannot be reduced any further due toWiFi’s CSMA, we exploit an additional dimension—redu-cing IL power consumption—in order to minimize itsenergy cost. Ideally, if the exact idle period is known, theradio could be powered off or put to sleep during IL, andwake up and process packets on demand. However, due tothe distributed nature of CSMA, the idle time betweenpackets varies widely and unpredictably. Underestimationof an idle interval will waste energy, while an over-estimation causes the radio to drop all incoming packetsduring the sleep.

So, one may raise an important question: “is it possible toput the radio in a subconscious mode, where it consumeslittle power and can still respond to incoming packetspromptly?” We answer this question by proposing Energy-Minimizing idle Listening (E-MiLi) that reduces the clockrate of the radio during its IL period. The powerconsumption of digital devices is known to be proportionalto their voltage square and clock rate [7], [8]. Theoretically,by reducing clock rate alone, E-MiLi reduces the IL’spower consumption linearly.

IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012 1441

. The authors are with the Real-Time Computing Laboratory, Department ofElectrical Engineering and Computer Science, The University of Michigan,2260 Hayward St., Ann Arbor, MI 48109-2121.E-mail: {xyzhang, kgshin}@eecs.umich.edu.

Manuscript received 12 Nov. 2011; revised 6 Feb. 2012; accepted 19 Apr.2012; published online 1 May 2012.For information on obtaining reprints of this article, please send e-mail to:[email protected] and reference IEEECS Log Number TMCSI-2011-11-0607.Digital Object Identifier no. 10.1109/TMC.2012.112.

1536-1233/12/$31.00 ! 2012 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

Page 2: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

It is, however, nontrivial to ensure that packets can bereceived at a lower clock rate than required. To decode apacket, the receiver’s sampling clock rate needs to be atleast twice the bandwidth of the transmitted signal,following the Nyquist’s Theorem. WiFi radios have alreadybeen optimized under this theorem by matching thereceiver’s clock rate with the Nyquist rate.

E-MiLi meets this challenge via a novel approach calledSampling Rate Invariant Detection (SRID). SRID separates thedetection from the decoding of a packet. It adds a specialpreamble to each 802.11 packet, and incorporates a linear-time algorithm that can accurately detect the preamble evenif the receiver’s clock rate is much lower than thetransmitter’s. SRID embeds the destination address into thepreamble, so that a receiver may only respond to packetsdestined for it. Upon detecting this special preamble, thereceiver immediately switches to the full clock rate and thenrecovers the packet with a legacy 802.11 decoder.

E-MiLi allows SRID to be integrated into existing MACor sleeping-scheduling protocols, using a simple Opportu-nistic Downclocking (ODoc) scheme. ODoc enables fine-grained, packet-level power management by adding adownclocked IL (dIL) mode into the radio’s state machine.ODoc exploits the burstiness and correlation structure ofreal traffic to assess the potential benefit of downclocking,and then downclocks the radio only if it is unlikely to incursignificant overhead.

We have implemented an E-MiLi prototype on theGNURadio/USRP platform [9]. Our experimental evalua-tion shows that E-MiLi can detect packets with close to100 percent accuracy even if the radio operates at 1

16 of thenormal clock rate. Within a normal SNR range (>8 dB),E-MiLi performs comparably to a legacy 802.11 detector.Furthermore, from real traffic traces, we find that for themajority of clients, the overall energy saving with E-MiLi

is close to that in pure IL mode with the maximumdownclocking factor. According to our measurements, thiscorresponds to 47.5 percent for a typical WiFi card with adownclocking factor of 4, and 36.3 percent for a softwareradio with a downclocking factor of 8. Further, ourpacket-level simulation results show that E-MiLi reducesenergy consumption consistently across different trafficpatterns, without any noticeable performance degradation.

In summary, thispapermakes the followingcontributions.

. Exploration of the feasibility and cost of fine-grainedcontrol of radio clock rate to improve energyefficiency.

. Design of SRID, a novel packet detection algorithmthat makes it possible to detect packets even if thereceivers are down clocked significantly.

. Introduction of ODoc, a generic approach to inte-grating SRID with existing MAC- and sleep-schedul-ing protocols.

. Implementation of E-MiLi on a software radioplatform and validation of its performance with realtraces and synthetic traffic.

The remainder of this paper is organized as follows:Section 2 analyzes the energy cost of IL in WiFi networksand describes the motivation behind E-MiLi. Section 3presents a measurement study of the relation between

energy consumption and clock rate in WiFi and softwareradio devices. Following an overview of E-MiLi (Section 4),Sections 5 and 6 present the detailed design of SRID andODoc, respectively. Section 7 evaluates E-MiLi. Section 9reviews related work and Section 10 concludes the paper.

2 WHY E-MILI?

In this section, we motivate E-MiLi by showing a largefraction of time and energy spent in IL for real-world WiFiusers. We also briefly discuss the reasons for the highpower-consumption of IL by anatomizing a typical radio.

2.1 Cost of Idle Listening

We acquired packet-level WiFi traces from publicly avail-able data sets: SIGCOMM’08 [10] and PDX-Powell [11]. Theformer was collected from a WLAN used for a conferencesession that has a peak (average) of 31 (7) clients. The latterwas collected from a public hotspot at a university book-store, with a peak (average) of 7 (3) clients. We built asimulator that can parse the traces and compute eachclient’s sojourn time in different states, including:

. TX&RX. The client is transmitting or receiving apacket.

. Sleep. The client is put to sleep. A client sets thepower-management field in its packet header to 1 ifit intends to sleep after the current frame transmis-sion and ACK [5].

. Idle listening. A state other than the above two. Thisincludes sensing the channel, waiting for incomingpackets, receiving packets not addressed to it, etc.We exclude the SIFS time, which is a short interval(9! 20 !s [5]) between two immediate packets (e.g.,in between data/ACK). We also consider a clientdisconnected if it does not transmit/receive anyunicast packets for 5 minutes or longer.

Fig. 1a plots the normalized fraction of time spent in thethree modes, distributed among all the clients in theSIGCOMM ’08 trace. More than 90 percent of clients enablepower management and judiciously put their radios tosleep. However, clients spend most of the time in IL, ratherthan sleeping: the median IL time is 0.87, and is above 0.6for more than 80 percent of clients. One may guess thereason for this to be the excessive contention in this busynetwork. However, even in the PDX-Powell trace (Fig. 1b),the IL time exceeds 0.52 for more than 70 percent of clients.In contrast, the actual TX&RX time is below 0.1 for morethan 90 percent of clients in both networks. Since WiFi’sPSM cannot eliminate MAC-layer contention and queuing

1442 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012

Fig. 1. CDF of the fraction of time spent in different modes for(a) SIGCOMM ’08 trace and (b) PDX-Powell trace.

Page 3: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

delays [6], the IL still dominates the TX&RX time by asignificant margin.

We further analyze the energy cost of IL. Since informa-tion on the actual type of clients’ WiFi cards is unavailable,we assume that their energy profile follows that of a typicalAtheros card [12, Section 10.1.5] (TX: 127 mW, RX:223.2 mW, IL: 219.6 mW, Sleep: 10.8 mW). Although theirabsolute power consumption differs, many widely usedWiFi cards have consistent relative power consumptionamong different states [13]. Fig. 2a shows that in a busynetwork, for more than 92 percent of clients, 90 percent ofenergy is spent in IL, i.e., IL costs nine times more energythan TX&RX for most clients. Moreover, although the sleeptime is substantial, the sleep power is negligible, whereasthe IL power is comparable to the TX/RX power, so themajority of cost is still with IL. For a network with lesscontention (Fig. 2b), IL costs less, yet still accounts for morethan 73 percent of energy cost for 90 percent of clients. Notethat the sleep energy may exceed the TX&RX energy, due tothe significant amount of sleep time.

The above evaluation reveals that IL accounts for themajority of a WiFi radio’s energy cost, and optimizing theIL time alone using PSM is not enough. If the IL power canbe reduced, it will clearly improve the energy efficiency ofPSM-like sleep scheduling protocols. In addition, for real-time applications, the constant active mode (CAM) of WiFiis preferable, since PSM may incur an excessive delay anddegrade the QoS [2]. By reducing IL power, even CAM canachieve high energy efficiency.

2.2 Why Is Idle Listening So Costly?

Intuitively, a radio should consume less power when it isnot actively decoding or transmitting packets, but the ILpower of commodity WiFi and other carrier-sensingwireless (e.g., ZigBee) devices is comparable to their TX&RXpower [2], [12], [14]. In what follows, we briefly discuss thereason for this by anatomizing the radio hardware.

Fig. 3 illustrates the architecture of a typical WiFireceiver (based on an Atheros 802.11 chip [15]). Anincoming signal is first passed through the RF and analogcircuit, amplified and converted from RF (e.g., 2.4 GHz) tothe baseband by a mixer. The analog baseband signal issampled by an Analog-to-Digital Converter (ADC), and theresulting discrete samples are passed to the CPU (basebandand MAC processor), which decodes the signal and re-covers the original bits in the data frame. The entire radio isdriven by a 40 MHz crystal oscillator, which feeds twopaths. The first is the frequency synthesizer that generatesthe center frequency used for the RF and analog mixer. Theother is the Phase-Locked-Loop (PLL) that generates the

clocking signal for the digital circuit: the sampling clock forthe ADC, as well as the main clock for the CPU.

Existing studies have shown the ADC and CPU to be themost power-hungry components of a receiver. In theAtheros 5001X chipset, for example, they account for55.3 percent of the entire receiver power budget [16,Table 5]. ADC and CPU power consumptions are alsosimilar (1.04:1 [17]). During IL, both the analog circuits andthe ADC operate at full workload as in the receiving mode.Moreover, the decoding load of the CPU is alleviated, but itcannot be put into sleep—it needs to operate at full clockrate in order to perform carrier sensing and packetdetection. This is the reason why IL power consumptionis comparable to that of receiving packets.

A similar line of reasoning applies to other wirelesstransceivers such as software radios. In software radios, theADC feeds the discrete samples to an FPGA, which mayfurther decimate (downsample) the samples and then sendthem to a general processor that serves as the basebandCPU. The similarity in hardware components implies thatsoftware radios are likely to suffer from the same problemwith IL. Considering the trend of software radios gettinggradually integrated into mobile platforms to reduce thearea cost [18], it is imperative to incorporate a mechanism toreduce its IL power.

3 IL POWER VERSUS CLOCK RATE

We propose to reduce the IL power by slowing down theclock that drives the digital circuitry in a radio. Moderndigital circuits dissipate power when switching betweenlogic levels, and their power consumption follows P / V 2

ddf ,where Vdd is the supply voltage and f the clock rate [7], [8].Hence, a linear power reduction can be achieved byreducing clock rate. In practice, due to the analogperipherals, the actual reduction is less than ideal. Forexample, in the ADC used by an Atheros WiFi chip [19],halving the sampling clock rate results in a 31.4 percentpower reduction. Here, using detailed measurements, weverify the actual effects of reducing the clock rate for bothWiFi NIC and the USRP software radio.

3.1 WiFi Radio

According to IEEE 802.11-2007 [5], the OFDM-based PHYsupports 2 downclocked operations with 10 MHz (half-clocked) and 5 MHz (quarter-clocked) sampling rate, inaddition to thedefault full-clocked20MHzoperation.We testthese twomodes on the LinkSysWPC55AGNIC (version 1.3,Atheros 5414 chipset), with a development version ofMadwifi (trunk-r4132), which supports eight half-clockedand 18 quarter-clocked channels at the 5 GHz band. Thedownclocked modes can be enabled by activating the “USAwith 1

2 and 14 width channels” regulatory domain on the NIC.

ZHANG AND SHIN: E-MILI: ENERGY-MINIMIZING IDLE LISTENING IN WIRELESS NETWORKS 1443

Fig. 2. CDF of the fraction of energy spent in different modes for

(a) SIGCOMM 2008 trace and (b) PDX-Powell trace.

Fig. 3. Architecture of a WiFi receiver.

Page 4: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

As to measurement of the WiFi’s power consumption,our approach is similar to that in [13]. We attach the NIC toa laptop (Dell 5410) powered with an external AC adapter,and use a passive current probe (HP1146A) and voltageprobe (HP1160) together with a 1 Gsps oscilloscope(Agilent 54815A) to measure the power draw. The actualpower consumption is the difference between the measuredpower level in different radio modes and the base levelwith the NIC removed. During the measurement, we tunethe WiFi to a channel unused by ambient networks. The ILpower is measured when the NIC is activated but nottransmitting/receiving packets. The TX/RX power ismeasured when the WiFi is sending/receiving one-wayping-broadcast packets at the maximum rate (100 packetsper second). The different clock modes are configured touse the same bit rate (6 Mbps) and packet size (1 KB).Table 1 shows the measurement results.

It can be seen that the power consumption decreasesmonotonically with clock rate. In particular, compared toa full-clocked radio, the IL power is reduced by 36 and47.5 percent for half-clocked and quarter-clocked mode,respectively. The absolute reduction is found differentfrom that reported in an existing measurement study [3].We guess this discrepancy results from the use of adifferent WiFi card (Atheros 5212) in their experiment. Asvalidated in [3], different NICs have very different powerprofiles at different clock rates. To confirm that the powerconsumption versus clock-rate relation is not limited tothe WiFi radio, we have also conducted experiments withthe USRP software radio.

3.2 Software RadioThe original USRP is driven by an internal 64 MHz clock,which is used by both the ADC and FPGA. We enabled theexternal clocking feature by resoldering the main clockcircuit, following the instructions in [9]. We use the USRPE100 [9] as an external clock source, which has aprogrammable clock generator (AD9522) that producesreference clocks below 64 MHz.1

We mounted an XCVR2450 daughter board on the USRP,which was then connected to the PC host (a Dell E5410laptop). The IL mode runs the standard 802.11a/g carriersensing and packet detection algorithm (see Section 7 forthe details of our implementation). The TX mode sends acontinuous stream of samples prepended with 802.11preambles. Since a complete 802.11 decoding module isunavailable, we only measure the IL and TX power. Wemeasure the USRP power directly with the oscilloscope andcurrent/voltage probes, and then add the power consump-tion of the external clock [20], which is 0.55 W and does not

vary with clock rates. Note that the normal clock rate ofUSRP is 64 MHz, whereas the maximum signal bandwidthsent to the PC is 4 MHz since the FPGA downsamples(decimates) the signals. While reducing the clock rate, weensure the signal bandwidth is decreased by the same ratioby adjusting the decimation rate.

Table 2 shows the measurement results. Similar to a WiFiradio, the USRP power consumption decreases monotoni-cally with clock rate. A power reduction of 22.5 percent(36.3 percent) is achieved for a downclocking factor of 2 (8).We found that at a 4 MHz clock rate (a downclocking factorof 16), the USRP can no longer be tuned to the 2.4 GHzcenter frequency, but the ADC can still be tuned correctly to4 MHz sampling rate, and power consumption decreasesfurther.

Since the PC host consumes a negligible amount of powerwhen processing the 4 MHz signal, we have omitted itspower consumption in Table 2. Future mobile softwareradio systems may incorporate dedicated processors toprocess the baseband signals. By reducing the processors’clock rate in parallel with the ADC and FPGA, the entiresoftware radio platform can achieve higher energy efficiency.

4 AN OVERVIEW OF E-MILI

E-MiLi controls the radio clock rate on a fine-grained, per-packet basis, in order to reduce the energy consumption ofIL. It opportunistically downclocks the radio during IL, andthen restores it to full clock rate before transmitting or afterdetecting a packet. Fig. 4 illustrates the flow of coreoperations when E-MiLi receives and transmits packets.

E-MiLi prepends to each 802.11 packet an additionalpreamble, called M-preamble. During its IL period, adownclocked receiver continuously senses the channeland looks for the M-preamble, using the SRID algorithm.Upon detecting an M-preamble, the receiver immediatelyswitches back to full clock rate, and calls the legacy 802.11decoder to recover the packet. The receiver leverages animplicit, PHY-layer addressing mechanism in SRID to filter

1444 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012

TABLE 1Mean Power Consumption (in W)

of WiFi under Different Clock Rates

TABLE 2Mean Power Consumption (in W)

of USRP under Different Clock Rates

Fig. 4. Idle listening and RX/TX operations in E-MiLi.

1. The USRP E100 cannot be tuned to signals below 32 MHz. So, we useda signal generator to produce clock signals below 32 MHz, with the sameconfiguration as those produced by the E100.

Page 5: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

the M-preamble intended for other nodes, and henceprevents unnecessary switching of clock rate.

A TX operations follow the legacy 802.11 MAC, exceptthat the carrier sensing is done by SRID. If the radio isdownclocked during carrier sensing and backoff, it needs torestore full clock rate before the actual transmission. Theexact restoration time is scheduled by another componentof E-MiLi, called Opportunistic Downclocking.

After completing an RX or TX operation, the radio cannotdownclock greedily. As we will verify experimentally inSection 6, switching clock rate takes 9.5 to 151 !s for atypical WiFi radio. During the switching, the clock isunstable, and packets cannot be detected even with SRID.To reduce the risk of packet loss, E-MiLi employs ODocagain to make a downclocking decision using a simpleoutage-prediction algorithm, which estimates if a packet islikely to arrive during the clock-rate switching.

In addition, after sending the M-preamble, a transmittercannot wait silently during the receiver’s switching period;it may otherwise lose the medium access and be preemptedby other transmitters. To compensate for the switching gap,the transmitter inserts a sequence of dummy bits betweenthe M-preamble and the 802.11 packet. The dummy bitscover the maximum switching period so that the channel isoccupied continuously. Note that the transmitter alwayssends the M-preamble, dummy bits, and 802.11 packets atthe full clock rate. It need not know the current clock rate ofthe receiver.

When multiple clients coexist, E-MiLi assigns a broad-cast address as well as multiple unicast addresses, eachwith a unique feature. This feature is embedded in the M-preamble and detectable only by the intended receiver. Toreduce the overhead of M-preamble, E-MiLi incorporatesan optimization framework that allows multiple clients toshare addresses at minimum cost.

In summary, E-MiLi always runs at full clock rate totransmit or decode packets, but downclocks the radio duringIL to detect implicitly addressed packets, whenever possible.Next, we detail the design of components in E-MiLi.

5 SAMPLE RATE INVARIANT DETECTION

To realize E-MiLi, its packet-detection algorithm mustovercome the following challenges: 1) it must be resilient tothe change of sampling clock rate; 2) it must be able todecode the address information directly at low samplingrates; and 3) due to unpredictable channel condition andnode mobility, its decision rule should not be tuned atruntime, and hence must be resilient against the variation ofSNR. We propose SRID to meet these challenges via a jointdesign of preamble construction and detection.

5.1 Construction of the M-Preamble

E-MiLi constructs the M-preamble to facilitate robust,sampling-rate invariant packet detection, while implicitlydelivering the address information. An M-preamble com-prises C"C # 2$ duplicated versions of a pseudorandomsequence, as shown in Fig. 5 (where C % 3).

Within the M-preamble duration, the channel remainsrelatively stable, and therefore the duplicated sequencessent by the transmitter maintain strong similarity at the

receiver. Hence, a receiver can exploit the strong self-correlation between the C consecutive sequences to detectthe M-preamble. More importantly, since radios samplesignals at a constant rate, the receiver would obtain Csimilar sequences even if it down samples the M-preamble.

To enhance resilience to noise, the random sequence inM-preamble must have a strong self-correlation property—itshould produce the best correlation output only whencorrelating with itself. The Gold sequence [21] satisfies thisrequirement. It outputs a peak magnitude only for perfectlyaligned self-correlation, and correlating with any shiftedversion of itself results in a low, bounded magnitude. For aGold sequence of length L % 2l ! 1 (l is an integer), the ratiobetween the magnitude of self-correlation peak and thesecondary peak is at least 2

l!12 . The original Gold sequence is

binary [21]. To make it amenable for WiFi transceivers, weconstruct a complex Gold sequence (CGS), in which the realand imaginary parts are shifted versions of the same Goldsequence generated by the standard approach [21].

In addition, we use the length of the CGS to implicitlyconvey address information. An address is an integernumber n, and corresponds to a CGS of length "TB & nDm$,whereDm is the maximum downclocking factor of the radiohardware. TB is the minimum length of the CGS used for thepreamble, also referred to as base length. To detect its ownaddress (e.g., n), at each sampling point t, the client simplyself-correlates the latest TB samples with the previous TB

samples offset by nDm. When the client is downclocked by afactor ofD, it scales down the base length to TBD!1 and offsetto nDmD!1 accordingly. The nDm value ensures thatdifferent addresses are offset by at least 1 sample, even ifthe CGS is downsampled by the maximum factor Dm.

One challenge related to the Gold sequence is that it onlyallows length of L % 2l ! 1. Hence, not all of the "TB & nDm$samples can be exactly matched to a whole Gold sequence.We solve this problem by first generating a long CGS, andthen assign the subsequence of length "TB & nDm$ to thenth address.

Clearly, to meet its design objectives, an ideal randomsequence for M-preamble should have strong self-correla-tion even after it is downsampled and truncated (since we onlyuse TB of the TB & nDm samples to perform self-correlation).We conjecture there does not exist such a sequence unless thesequence length is very large and the downsampling factor issmall. We leave the theoretical investigation of this problemas our future work. In this paper, we will empirically verifythat the CGS with a reasonable length suffices to achievehigh detection accuracy in practical SNR ranges.

5.2 Detection of the Preamble

We formally derive the detection algorithm in SRID bymodeling how the receiver down nsamples the M-preambleand identifies it via self-correlation.

ZHANG AND SHIN: E-MILI: ENERGY-MINIMIZING IDLE LISTENING IN WIRELESS NETWORKS 1445

Fig. 5. M-preamble construction and integration with an 802.11 packet.

Page 6: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

Let T % C"TB & nDm$ be the total length of the M-preamble (Fig. 5), and x"t$; t 2 '0; T $, the transmittedsamples corresponding to the M-preamble. For a full-clocked receiver, the received signals are

yo"t$ % e2"!fth"t$x"t$ & n"t$; t 2 '0; T $; "1$

where n"t$ is the noise, h"t$ the channel attenuation (acomplex scalar representing amplitude and phase distor-tion), and !f the frequency offset between the transmitterand the receiver. When a receiver operates at the clock rateof 1

D (i.e., with a downclocking factor of D), the receivedsignals become

z"k$ % e2"!fth"t$x"t$ & n"t$; t % kD; 0 ( k <T

D

! ":

Here,Dmust be an integer divisor of the base length TB ofthe CGS, i.e., bTB

D c %TBD %4 T1. To detect M-preamble, at each

sampling point k, the receiver with address n performs self-correlation between the latest T1 samples and the previous T1

samples offset by nDmD!1, resulting in

R"k$ %Xk&T1!1

i%k

z"i$z)"i! T1 ! nDmD!1$ "2$

*Xk&T1!1

i%k

e2"!fiDh"iD$x"iD$#e2"!f"iD!TB!nDm$

h"iD! TB ! nDm$x"iD! TB ! nDm$$)

"3$

* eTB&nDm jh"kD$j2Xk&T1!1

i%k

jx"iD$j2; "4$

where "+$) denotes the complex conjugate operator.Equation (3) is derived based on the fact that the signal

level is usually much higher than the noise. Equation (4) isbased on the fact that 1) the random sequence x"t$preserves similarity with its predecessor sequence, eventhough it is downsampled; and 2) the channel remainsrelatively stable over its coherence time, which is muchlonger than the preamble duration. To see this, we note thatthe coherence time can be gauged as To % #%%

2p

"v, where #

and v denote the wavelength of the signal and the relativespeed between the transmitter and the receiver [22]. At awalking speed of 1 m/s, To equals 28.8 milliseconds,whereas the M-preamble duration lasts for tens of micro-seconds (see Section 5.3.1).

Meanwhile, the energy level of T1 samples is calculated as

E"k$ %Xk&T1!1

i%k

jz"i$j2 * jh"kD$j2Xk&T1!1

i%k

jx"iD$j2: "5$

From (4) and (5), we get jR"t$j * E"t$. By contrast, if noM-preamble presents or an M-preamble with a differentaddress a is transmitted, then the self-correlation yields

jR"k$j * jh"kD$j2&&&&&Xk&T1!1

i%k

x"iD$x"iD! TB ! aDm$)&&&&& * 0:

This is because the sequence x"iD$; i 2 'k; k& T1 ! 1, is atruncated CGS and has strong correlation only with itself.

Fig. 6 shows a snapshot of jR"t$j and E"t$when receivinga packet prepended with M-preamble. jR"t$j aligns almostperfectly with E"t$ in an M-preamble, even though thereceiver is downclocked. In contrast, jR"t$j differs from E"t$significantly if noise or uncorrelated signals are present.

Based on the above findings, SRIDuses the following basicdecision rule to determine the presence of an M-preamble:

H < jR"k$j + 'E"k$,!1 < H!1; "6$

where H is a threshold such that H *< 1. This decision rulehas several key advantages. First, it normalizes the self-correlation with the energy level, so H need not be changedaccording to the signal strength. We will show experimen-tally (Section 7) that a fixed value ofH % 0:9 is robust acrossa wide range of SNR. Second, it does not require estimationof the channel parameters or calibration of the frequencyoffset, and hence can be used in dynamic WLANs with userchurn and mobility.

For further enhancement of resilience to noise, note that

the decision rule (6) is likely to be satisfied at all the sampling

points from the second to the Cth CGS (Fig. 5). There are"C!1$"TB&nDm$

D %4 T2 such points at a downclocking factor D,

which can offer high diversity in a noisy or fading

environment. To exploit this advantage, at each sampling

point k, SRID stores the decision for the past T2 samples in a

FIFO queue, and then apply the following enhanced rule:

for k! T2 < i ( k, the number of sampling points satisfying

(6) #H1T2, whereH1 is a tolerance threshold andH1 2 "0; 1,.In addition, during idle periods (i.e., when no signal is

present), both the self-correlation and the energy level maybe close to 0 and close to each other, and hence the decisionrule (6) may be falsely triggered. To prevent such falsealarms, we added an SNR squelch, which maintains amoving average of incoming signals’ energy level, with thewindow size equal to T1:

Ea"k$ % T!11 E"k$ &

'1! T!1

1

(Ea"k! 1$: "7$

The SNR squelch passes a sampling point to the self-correlator only if its SNR exceeds a threshold Hs, whichcorresponds to the minimum detectable SNR (set to 4 dB forSRID). Since an idle period (noise floor) usually precedesthe M-preamble (with length TD!1) due to the MAC-layercontention, the SNR level can be estimated as

SNR % 10 log10Ea"t$

Ea"t! T $: "8$

1446 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012

Fig. 6. Detecting M-preamble using SRID (clock rate % 1=4).

Page 7: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

Algorithm 1 summarizes the detection of M-preamble inSRID. For each time stamp (sampling point), both the self-correlation in (2) and the energy level in (5) can becomputed by a single-step operation, which updates themetrics with an incoming signal and subtracts the obsoletesignal. Hence, the algorithm has linear complexity withrespect to the number of samples, and is well suited forimplementation on an actual baseband signal processor.

Algorithm 1. Detecting the M-preamble using SRID.1. Input: new sample z"k& T1 ! 1$ at sampling point

k& T1 ! 12. Output: packet detection decision at sampling point k3. =)Update energy level of past T1 samples)=4. E"k$ E"k! 1$ & jz"k& T1 ! 1$j2 ! jz"k! 1$j2

5. =)Update average energy level)=6. Ea"k$ T!1

1 E"k$ & "1! T!11 $Ea"k! 1$

7. =)Update self-correlation with predecessor sequence)=8. R"k$ R"k! 1$ & z"k& T1 ! 1$z"k! nDmD!1 ! 1$)

9. !z"k! 1$z"k! 1! T1 ! nDmD!1$)

10. =)Apply SNR squelch and self-correlation decision)=

11. if 10 log10Ea"k$

Ea"k!TD!1$ > Hs && H < jR"k$jE"k$ < H!1

12. then decisionQ push 113. else decisionQ push 014. fi15. if sum(decisionQ) > H1 + "C!1$"TB&nDm$

D16. then return 117. fi18. return 0

5.3 Address Allocation

5.3.1 Minimum-Cost Address Sharing

Since M-preamble uses sequence length to convey addressinformation, the addressing overhead increases linearlywith network size. For a network with N nodes, the M-preamble has a maximum length of C"TB &NDm$. In ourimplementation, the base length TB % 64, and CGS repeti-tion C % 3. For a medium-sized network, say N % 5, and amaximum downclocking factor Dm % 4, the entire M-preamble would have a length of 252. When transmittedat a 20 MHz sampling rate, the M-preamble only takes252

2-107 s % 12:6 !s channel time, which is comparable to the16 !s overhead of the 802.11a/g preamble [5]. However, fora large network, e.g., N % 50, the M-preamble overheadincreases to 69:6 !s, which may be overly large, especiallyfor short packets.

To reduce the addressing overhead, E-MiLi allowsmultiple clients to share a limited number of addresses.Address sharing, however, introduces side effects: clientsmay unnecessarily trigger each other, thus incurring extraenergy consumption. E-MiLi makes a tradeoff by carefullyallocating addresses according to clients’ relative channelusage, i.e., the ratio of each client’s TX&RX time to the totalTX&RX time of the WLAN. The intuition behind this is thata client that transmits/receives packets more frequentlyshould share his address with a fewer number of otherclients, so as to minimize the cost of sharing.

We formalize this intuition with an optimization frame-work. Given the number of clients N , and the maximum

address Km, we seek the optimal address allocation thatminimizes the overhead of E-MiLi, as follows:

minXKm

k%1

Lk

XN

i%1

piuik

!XN

i%1

uik

" #

"9$

s:t:XKm

k%1

uik % 1; 8i 2 '1; N ,; "10$

uik 2 f0; 1g; 8i 2 '1; N ,; 8k 2 '1; Km,; "11$

where Lk is the overhead when the address k is used. pi isclient i’s relative channel usage, and uik a binary variableindicating whether or not client i uses address k. Intuitively,the objective function (9) represents the sum of theoverhead of each address, weighted by sum of the channelusages of all clients sharing that address and furthermultiplied by the number of such clients. The multiplicationis necessary because a packet with address k triggers allclients with address k. Equation (10) enforces the constraintthat each client uses only one address.

This optimization problem is a nonlinear integer pro-gram, which is NP-hard in general. In our actual imple-mentation, we approximate the solution by relaxing theinteger constraint (11) to 0 ( uik ( 1, solving the resultingquadratic optimization program, and then rounding theresulting uik back to its integer value. To implement theaddress sharing algorithm, the AP needs to periodically(e.g., every 1 minute) compute the relative channel usage pi,and then broadcast the new allocation to all clients.

To test the effectiveness of the approximation, we runthe address sharing algorithm on the SIGCOMM ’08 trace(assuming Km % 5 and Lk % kDm) and plot the totaladdress overhead of E-MiLi in Fig. 7. We observe thatthe integer-rounding-based solution closely approximatesthe lower bound enforced by the quadratic optimizationover 0 ( uik ( 1. On average, the approximate solutionexceeds the lower bound by only 1.8 percent. Fig. 7 alsoshows the mean overhead of an algorithm that randomlyassigns an address for each client (error bar shows standarddeviation over 20 runs). We observe that the approximationalgorithm can save more than 50 percent of overhead overthe random allocation.

5.3.2 The Broadcast Address

In addition to the address designed for each node, E-MiLiassigns a broadcast address known to the AP and all clients.It corresponds to an M-preamble with address n % 0.Therefore, each node needs to maintain a self-correlator withoffset nDm % 0, in addition to the one with its own address.

ZHANG AND SHIN: E-MILI: ENERGY-MINIMIZING IDLE LISTENING IN WIRELESS NETWORKS 1447

Fig. 7. Performance of address sharing algorithms.

Page 8: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

For the carrier sensing purpose, a node also needs toidentify the existence of packets from other transmitters.Similar to the original 802.11, SRID can perform both energysensing and preamble detection. The former is achieved byfollowing (7). When downclocked by a factor of D, a nodecan only sense D!1 of the energy compared with a full-clocked receiver. Hence, it reduces the energy detectionthreshold to D!1 of the original. When preamble-basedcarrier sensing is necessary, it can be realized by prepend-ing an additional broadcast preamble. When this firstpreamble is detected, the node determines the channel tobe busy, and continues to track the energy level of the entirepacket. However, it will restore full clock rate only when itdetects a second preamble, which is either addressed to it or isanother broadcast preamble.

E-MiLi can coexist with 802.11a/g clients even in thepreamble detection mode. The 802.11a/g [5] employs self-correlation to detect a short preamble, which corresponds toa random sequence in the frequency domain, and a periodicsequence (period 16, with 10 repetitions) in the timedomain. It can be considered as a subset of SRID, withbase length TB % 16, sequence repetition C % 10, nodeaddress 0 and no downclocking, and thus can be easilydetected by E-MiLi clients. On the other hand, by replacingthe first preamble with an 802.11 preamble, E-MiLi nodescan be detected by legacy 802.11 as well.

6 OPPORTUNISTIC DOWNCLOCKING

We now present the ODoc module, which schedules thedownclocking to balance its overhead and maintaincompatibility with existing MAC and sleep schedulingprotocols. We start by inspecting the overhead in switchingclock-rates.

6.1 Delay in Switching Clock Rates

When switching to a new clock rate, the radio needs to bestabilized before transmitting/receiving signals. Since thefrequency synthesizer and analog circuit’s center frequencyremain the same, the time cost mainly comes fromstabilizing the digital PLL (driving the ADC and CPU).This is only several microseconds in state-of-the-art WiFiradios. For example, the PLL takes less than 8 !s to stablizeitself in the MAXIM 2831 RF transceiver [23, p. 17], and thedigital circuit (ADC and CPU) needs only 1:5 !s to reset [24,Section 7.1], so the total switching time is below 9:5 !s.

We have also measured the switching delay of theAtheros 5414 NIC. We modified the ath5k driver that candirectly access the hardware register and reset the clockrate. After changing the clock-rate register, we repeatedlycheck a baseband testing function until it returns 1 (aconventional way of verifying if the ADC and basebandprocessor have become ready to receive packets in ath5k),and then record the duration of this procedure.

According to our experimental results, switching betweenclock-rate 1 and 1

4 takes 139 to 151 !s, whereas switchingbetween 1 and 1

2 takes 120 to 128 !s. We note that this is aconservative estimation of the actual switching delay. Toswitch to a new rate, the Atheros NIC needs to reset not justthe PLL, but also all registers for the OFDM decoding andMAC blocks in the CPU, so that the entire receiver chain can

run a valid 802.11 mode. In contrast, E-MiLi only needs toreset thePLL,while keeping the registers in theCPU intact. Inaddition, the latency induced by the baseband testingfunction and its interface to the PC host is unknown, but isincluded in the switching delay in our measurement.

We will henceforth use the 9:5 !s switching delay for theMAXIM 2831 chip as a lower bound, and use themeasurement result for Atheros 5414 as an upper bound,although the ODoc module is not restricted to these bounds.

6.2 Scheduling of Downclocking

6.2.1 Control Flow

E-MiLi interacts with the WiFi MAC/PHY using a simpleinterface. On the one hand, WiFi calls E-MiLi (the SRIDmodule) to assess the channel availability. On the otherhand, E-MiLi obtains the radio’s state machine from theWiFi MAC and the sleep scheduler. Whenever the radiotransits to IL, E-MiLi calls its ODoc module to determinewhether and when to switch clock rate.

Fig. 8 illustrates the state machine of E-MiLi. In dILmode, the radio runs SRID continuously, and switches tothe full-clocked RX mode immediately upon detection of anM-preamble. When there are packets to be transmitted,carrier sensing is performed by SRID, but the MACschedule strictly follows the 802.11 CSMA/CA algorithm.ODoc continuously queries the 802.11 backoff counter, andreverts the radio to full clock rate when the countdownvalue of the backoff counter is less than Tc & SIFS, whereTc is the maximum switching delay, and SIFS is the shortinterframe space defined in 802.11 [5]. ODoc mandates theradio to perform carrier sensing within this SIFS intervalafter switching to full-clock rate, in order to ensure thechannel remains idle after switching. Otherwise, it needs tocontinue carrier sensing and backoff according to 802.11.

The state-transitions TX$ Sleep and RX$ Sleep aremanaged by 802.11 or other sleep-scheduling protocols.Whenever a TX or RX completes and the radio is not put tosleep, ODoc decides whether to switch to dIL or the normalIL mode. It makes this decision using an outage predictionscheme, as detailed next.

6.2.2 Outage Prediction

ODoc’s outage prediction mechanism decides if the nextpacket is likely to arrive before the radio is stabilized to anew clock rate (referred to as an outage event). It first checksif there will be a deterministic operation, i.e., an immediateresponse of the previous operation. For example, CTS,DATA, and ACK packets are all deterministic operations tofollow an RTS. Such packets are separated only by an SIFS,which is usually shorter than or comparable to the switch-ing time, so the radio must remain at full rate in between.

1448 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012

Fig. 8. Radio state-transition when integrating E-MiLi with 802.11.

Page 9: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

When a series of deterministic operations end, ODocchecks if an outage occurred recently. It maintains a binaryhistory for each nondeterministic packet arrival, with “1”representing that the interpacket interval is shorter than Tc,and “0” otherwise. It asserts that an outage is likely to occurand remains at full clock rate, if the recent history contains a“1.” The key intuition lies in the burstiness of WiFi traffic—ashort interval implies an ongoing transmission of certaindata, and is likely to continue multiple short intervals untilthe transmission completes.

An important parameter in ODoc is the size of history. Alarge history size may predict an outage when it does notoccur, thus missing an opportunity of saving energy bydownclocking. On the other hand, a small history size resultsin frequent misdetection of packets arriving within Tc.Fortunately, a misdetection causes only one more retransmission,because a missed packet will be detected in its nextretransmission, when the receiver has already been stabi-lized. Therefore, a small history size is always preferredwhen energy efficiency is of high priority. Aswill be clarifiedin our experimental study, a history size of between 1 and 10is sufficient to balance the tradeoff between false predictionand misdetection.

7 EVALUATION

In this section, we present a detailed experimental evalua-tion of E-MiLi. Our experiments center around twoquestions: 1) How accurate can E-MiLi detect packets ina real wireless environment, and with different down-clocking rates? 2) Howmuch of energy can E-MiLi save forreal-world WiFi devices and at what cost?

To answer these questions,wehave implementedE-MiLi

on software radios and network-level simulators as follows:

. We have implemented the SRID algorithm, includ-ing the M-preamble construction and detection, onthe GNURadio platform and verify it on a USRPtestbed. As a performance benchmark, we have alsoimplemented the 802.11 OFDM preamble encoding/detection algorithm (Section 5.3.2).

. E-MiLi’s energy efficiency depends on the relativetime of IL, which, in turn, depends on network delayand contention, and hence, we leverage real WiFitraces again to evaluate the energy efficiency of E-MiLi. We implemented the ODoc framework andaddress allocation algorithm by extending the trace-based simulator (Section 3), and then integratingresults from the SRID experiments.

. We have also implemented ODoc in ns-2.34, whichcan be used to verify the performance of E-MiLi

with synthetic traffic patterns (e.g., HTTP and FTP)independently.

7.1 Packet-Detection Performance

We test the detection performance of SRID under differentSNR levels and downclocking factors. The SNR is estimatedas SNR % Es!EN

EN, where Es is the average energy level of

incoming samples when a packet is present, and EN is thenoise floor, both smoothed using a moving average with thewindow size equal to the length of the M-preamble. Note

that this SNR value overestimates the actual SNR experi-enced by the decoder, since the decoding modules will raisethe noise level by around 3.5 dB [12]. Given that 802.11needs at least 9.7 dB SNR to decode packets [17], SRID mustbe able to detect packets accurately above 9.7 dB SNR.

We set the base length of SRID’s CGS to TB % 64, andmaximum downclocking factor Dm % 16. We fix the self-correlation threshold H % 0:9, and the tolerance thresholdH1 % 0:6 (Section 5). We will show that these thresholds arerobust across different experiment settings.

7.1.1 Single Link

We first test SRID on a single link consisting of two USRPnodes within Line-of-Sight (LOS). We downclock thereceiver by different factors, and vary the link’s SNR byadjusting the transmit power and link length/distance.Since the USRP fails to work when the external clock isdownclocked to 1

16 , we scale its FPGA decimation rate by 16,which is equivalent to downsampling the signals by a factorof 16. Under each SNR/clock-rate setting, the transmittersends 106 packets at full clock rate with constant interarrivaltime. The misdetection probability (Pm) is calculated by thefraction of time stamps where a packet is expected to arrivebut fails to be detected, and vice versa, for the false-alarmprobability (Pf ).

Fig. 9 plots Pm and Pf as a function of a link’s time-averaged SNR (rounded to integer values). Pm dropssharply as SNR increases, and approaches 0 as SNR growsabove 8 dB. It tends to be higher under a high downclockingfactor, mainly because fewer sampling points are availablethat satisfy the decision rule (6) and thus, SRID is moresusceptible to noise. When SNR % 4 dB and D % 16, Pm

grows up to 6 percent. Under practical SNR ranges (above9.7 dB), however, Pm is consistently below 1 percent for allthe clock rates. In addition, SRID shows a comparabledetection performance with 802.11. In fact, it may havelower Pm when the down-clocking factor D is below 16.This is because SRID uses a longer self-correlation sequencethan 802.11 (64 versus 16), which increases its robustness tonoise. The false-alarm probability Pf in Fig. 9b shows atrend similar to Pm.

Recall SRID uses nDm, the spacing between repetitiveCGS to convey address n. A natural question is: how largecan n be to ensure a high detection accuracy? Fig. 10 plotsthe detection performance as n increases. For a stationarylink, both Pm and Pf remain relatively stable. This isbecause even for the address n % 100, two self-correlationsequences are separated by 1,600 samples, corresponding to400 !s at the 4 MHz signal bandwidth of USRP, which iswell below the channel’s coherence time. For a mobile client

ZHANG AND SHIN: E-MILI: ENERGY-MINIMIZING IDLE LISTENING IN WIRELESS NETWORKS 1449

Fig. 9. SRID performance for a single link.

Page 10: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

(created by moving the USRP receiver around the trans-mitter at walking speed), the detection performance is onlyslightly affected by the address length, since the lowmobility causes SNR variations, but does not change thecoherence time significantly.

7.1.2 Testbed

We proceed to evaluate SRID on a testbed consisting of nineUSRP2 nodes (one AP and eight clients) deployed in alaboratory environment with metal/wood shelves and glasswalls. Fig. 11 shows a map of the node locations. Node D ismoving between point D and E at walking speed, and allothers are stationary. This testbed enables the evaluation ofSRID in a real wireless environment subject to effects ofmultipath fading, mobility, and NLOS obstruction. Moreimportantly, it allows testing the false-alarm rate due tocross correlation between different node addresses.

Due to the limited number of external clocks, we createthe effect of downclocking by changing the USRP2’sdecimation rate, so that the receiver’s sampling rate becomes1 to 1

16 of the transmitter’s. We allow the AP to send106 packets to each client in sequence. Fig. 12a shows that,depending on node locations, Pm varies greatly. In general,nodes farther away (e.g., H) or obstructed by walls (e.g., F )from the AP has higher Pm. The mobile node D may havehigher Pm than a node farther from the AP but is stationary(e.g., nodeE). Consistent with the single link experiment, thedownclocking factor 4 results in comparable Pm with 802.11.

Fig. 12b shows the false-alarm probability due to crosscorrelation, i.e., the probability that a client detects packetsaddressed to others. The relative Pf for different clientsshows a similar trend as Pm, depending on the location andmobility. Unlike the single link case, the Pf tends to belarger than Pm, because the cross correlation betweensequences has stronger effects on Pf than pure noise.Remarkably, even for the worst link and with D % 16, Pf isbelow 0.04, implying negligible energy cost due to falsetriggering. We note that for 802.11, the address field must bedecoded from the packet, so Pf here is not meaningful for it.

From the above experiments, we observe that SRID hasclose to 100 percent detection accuracy (and is comparable to

802.11) under practical SNR ranges and with downclockingrate up to 16. Hence, it can be used to realize E-MiLi inpractical wireless networks.

7.2 Improving WiFi Energy Efficiency

7.2.1 Real WiFi Traffic

We now evaluate E-MiLi’s energy efficiency through trace-based simulation. We obtain WiFi and USRP power-consumption statistics from actualmeasurements (Section 3).We use the 151 !s switching time of the Atheros AR5414NIC as the worst case estimate of switching delay, assumingthe power consumption during clock switching is the sameas in full-clocked mode. As we will clarify, an outage due tothe switching delay occurs with a less than 4.2 percentprobability, so we assume an outage event does not affectthe WiFi traces except causing one retransmission. Inaddition, we adopt the Pm and Pf values at 8 dB as aconservative estimation of the packet loss or false alarmcaused by SRID. Unless mentioned otherwise, 15 addressesare allocated and shared among all clients, and a history sizeof 5 is used in ODoc.

Energy savings. Fig. 13a illustrates the energy saving ofE-MiLi, assuming clients are using WiFi devices with amaximum downclocking factor of 4. For a large network(SIGCOMM ’08 traces [10]), the energy saving ranges from 41to 47.3 percent. Its CDF is densely concentrated—for around92 percent of clients, the energy saving ranges between 44and 47.2 percent, which is close to the 47.5 percent energysaving when a client remains in downclocked IL mode(Section 3). In a small network (PDX-Powell traces [11]) withless contention, IL induces less energy cost, so the energy-saving ratio of E-MiLi is relatively low. However, since ILtime still dominates, the median saving remains around44 percent, and minimum 37.2 percent. Fig. 13b plots theresults assuming clients’ power consumption is the same asthe USRP device with a maximum downclocking factor of 8.Again, the energy saving is concentrated near 36.3 percent,the saving in pure IL mode (Section 3).

1450 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012

Fig. 10. Detection performance versus the number of unique addresses.

Fig. 11. Network topology for evaluating SRID in a testbed.

Fig. 12. SRID performance in a USRP testbed.

Fig. 13. Energy saving ratio for (a) WiFi, maximum downclocking factorof 4; (b) USRP, maximum downclocking factor of 8.

Page 11: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

These experiments reveal that E-MiLi can explore themajority of IL intervals to perform downclocking. Itsenergy-saving ratio can be roughly estimated as $ % $cPIL,where $c is the energy-savings ratio in pure IL mode usingthe maximum downclocking factor, and PIL the percentageof idle listening energy during a radio’s lifetime. Since PIL isclose to 1 for most clients, $ is close to $c.

Overhead of E-MiLi and effect of ODoc. The overheadof E-MiLi comes from misdetection (and retransmission)due to a packet arriving in between the switching time.Such events can be alleviated by ODoc’s history-basedoutage prediction mechanism. In this experiment, weevaluate the cost of such outage and the effectiveness ofODoc in alleviating it. Fig. 14a shows that when history sizeequals 1, 4.2 percent packets may need to be retransmittedfor some clients. With a history size of 10, retransmission isreduced to below 0.8 percent for 90 percent of clients. Afurther increase of the history size to 100 shows only amarginal improvement. On the other hand, Fig. 14b shows asmall history size results in higher energy efficiency,implying that the energy savings from aggressive down-clocking dwarfs the small waste due to retransmissions.Hence, a small history size is preferable for ODoc if energyefficiency is of high priority.

7.2.2 Synthetic Traffic PatternsTo further understand E-MiLi’s benefits and cost undercontrollable network conditions, we implement and test itin ns-2.34. We compare performance of the legacy WiFi(including both CAM and PSM), and E-MiLi-enhancedWiFi (referred to as CAM& E-MiLi and PSM& E-MiLi). Wemodified the PHY/MAC parameters of ns-2 to be consistentwith that in 802.11g, and fix the data rate to 6 Mbps. Weimplement the ODoc based on 802.11, and configure it in asimilar manner to the trace-driven simulator. The PSMmodule builds on the 802.11 PSM extension to ns-2 [25], andthe power consumption statistics follow our measurementof AR5414 (Section 3). We evaluate two applications: webbrowsing and FTP, which have different performanceconstraints.

Web browsing.We simulate a web-browsing applicationusing the PackMIME http traffic generator in ns-2, whichprovides realistic stochastic models of HTTP flows. Thenetwork consists of one HTTP server connecting to a WLANAP via an ADSL2 link, with 1.5 Mbps (0.5 Mbps) downlink(uplink) bandwidth and exponentially distributed delaywithmean 15ms. TheAP serves oneHTTP client (withmeanpage request interval of 30 s) and multiple backgroundclients. Similar to [6], we study the effect of backgroundtraffic by running fixed-rate (200 Kbps, 512-byte packet size)UDP file transfer between theAPand the background clients.

Fig. 15a shows the energy usage of a 5-minute web-browsing session. PSM shows around 18 percent energysaving over CAM.CAM+E-MiLi saves 39.8 percent of energyover CAM without background traffic, and 47.1 percentwhen the number of background clients grows to 10. SincePSM optimizes the sleep schedule of clients, the ratio of ILtime is less, compared to CAM, and thus PSM& E-MiLiachieves less energy saving (33 to 37.1 percent) thanCAM& E-MiLi. Also, note that E-MiLi is relatively insensi-tive to background traffic, as it can enforce address filteringeven at low clock rate.

Fig. 15b plots the average per-page delay during the web-browsing session. Clearly, E-MiLi incurs a negligible delaywhen integrated into legacyWiFi. Although theM-preambleand clock switching costs channel time, it is much shorterthan the network and contention delay. Notably, PSM incursa longer delay than CAM due to its sleep schedulingmechanism, and CAM& E-MiLi has a shorter delay, yethigher energy efficiency than PSM.We expect an even betterenergy-delay tradeoff to be achieved by jointly designing thePSM sleep scheduling algorithm and E-MiLi. We leavesuch an optimization as our future work.

FTP.Weproceed to evaluate E-MiLi using the FTP trafficgenerator in ns-2, assuming a client downloads a 20 MB file(with packet size 1 KB) directly from the AP. Compared tothe fixed-duration web browsing, the FTP’s energy usage ismore sensitive to the background traffic (Fig. 16a), becausethe downloading duration is prolonged by MAC-layercontention. PSM is found to consume 36.8 to 39.4 percentmore energy than CAM, due to the fact that it may result inhigher energy-per-bit than CAM [1]. In addition, althoughE-MiLi achieves a similar level of energy saving as in theweb browsing, it may degrade the FTP throughput by up to4.4 percent in the absence of background traffic (Fig. 16b).This is due mainly to its overhead, i.e., the switching delay,the extra channel time of the M-preamble, and the imperfectdetector and outage predictor that incur MAC-layer retrans-missions. Moreover, note that we assume no end-to-enddelay and the throughput depends only onMAC contention,which zooms in the overhead from E-MiLi.

ZHANG AND SHIN: E-MILI: ENERGY-MINIMIZING IDLE LISTENING IN WIRELESS NETWORKS 1451

Fig. 14. Effects of history size (SIGCOMM ’08 trace). Fig. 15. Performance of a 5-minute web-browsing session.

Fig. 16. Performance when downloading a 20 MB file using FTP.

Page 12: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

Overhead in high-rate applications. One caveat toE-MiLi is that the duration of the M-preamble and theswitching delay are fixed, whereas the channel time fortransmission of useful data decreases as the data rateincreases. The overhead of E-MiLi will thus be amplifiedat a high data rate. We illustrate this effect by varying thePHY-layer data rate for an FTP file transfer with the numberof contending clients fixed at 6. Other parameter settings arethe same as in the FTP experiments above. The correspond-ing M-preamble takes 10:1 !s and the switching delay isagain configured to 151 !s. Fig. 17 shows that as the data rateincreases, CAM& E-MiLi causes CAM more throughputdegradation, and the amount of energy saving decreases dueto the longer time in transferring the data.When the data ratereaches 54Mbps,CAM& E-MiLi degrades the throughput ofCAM by 17.6 percent (the actual throughput is much lowerthan 54 Mbps due to inherent overhead and collisioninduced by TCP when running over 802.11 [26]), whilesaving 23.1 percent of energy. However, when takingadvantage of the short switching delay of recentWiFi chipset(e.g., 9:5 !s with MAXIM 2831), the throughput degradationis negligible, and the energy saving ratio is consistentlyaround 40 percent for all data rates. In addition, E-MiLi seesno throughput degradation when integrated with PSM, andthe resulting energy saving is kept around 30 percent.

Noted that the effect of fixed preamble overhead is aninherent problem of high data-rate 802.11 protocols, and canbe resolved by standard solutions such as the packetaggregation in 802.11n. Further, the effects of overhead ofE-MiLi becomes less severe in a busy network, wherecontention is high and the channel time consumed bypreamble and switching overhead becomes negligiblecompared to the contention delay. In addition, throughputis a critical metric only for rate-intensive applications likeFTP.Mobilewireless devices aremore likely to be dominatedby elastic traffic such as VoIP andHTTP. Such traffic patternstend to incur a significant amount of idle listening timewhichoverwhelms the channel time consumed by E-MiLi’spreamble and switching delay. As already exemplified inour web-\textbackslash browsing experiments, they canmake substantial energy saving by using E-MiLi.

8 DISCUSSION

Scalability to MIMO radios. The overhead of E-MiLi isfixed even if the NIC were equipped with a MIMO

transceiver. The overhead of E-MiLi mainly comes fromthe preamble and the clock switching delay. For MIMOsystems suchas 802.11n, all theRF chains of a receiverdetect asingle preamble embedded in each packet, and then usesdifferent preambles for channel estimation. Similarly, whenusing E-MiLi, they can share the same M-preamble forpacket detection. In addition, the clock switching delaydepends on the PLL settling time of each RF chain. ModernMIMO transceivers may either allow the RF chains to sharethe samePLL [27], or equip eachRF chainwith a separate PLL[28]. In the former case, the switching delay is fixed andshared amongall RF chains. In the latter case, the settling timeof all RF chains is similar and can overlap with each other.

In summary, neither the preamble overhead nor theswitching delay increases with the number of antennas in aMIMO system. Therefore, E-MiLi works for modernMIMO NICs without introducing any extra overheadcompared to the case of SISO NICs.

Enabling virtual carrier sensing.2 An E-MiLi receiveremploys SRID to detect packets intended for itself, and isable to carrier sense other packets via energy detection.However, energy sensing alone may not be enough toaddress a pathological case, i.e., the hidden terminalproblem. In IEEE 802.11, virtual carrier sensing is anoptional solution, which requires an RTS/CTS handshakebefore the actual data transmission. The RTS/CTS packetpiggy-backs a duration of the forthcoming data packet.Neighboring transmitters overhear the RTS/CTS and extendthe channel’s busy time by the corresponding duration.

In E-MiLi, virtual carrier sensing can be simply realizedas follows: A transmitter/receiver prepends RTS/CTS withthe broadcast preamble, so that all neighboring nodes candetect the RTS/CTS, restore full-clock rate and decode theduration field using a legacy 802.11 decoder. Then, as in the802.11 virtual carrier sensingmechanism, if the forth-comingdata packet is not intended for it, a node will enter the sleepmode and remain there throughout the packet duration.Since the data packets’ duration is usually much longer thanthe RTS/CTS, the energy consumption in decoding RTS/CTS is dominated by the energy savings with sleep, and thesavings in IL energy remain the same. Hence, with thissimple mechanism, E-MiLi will retain its advantages overlegacy approaches with virtual carrier sensing enabled.

Note that the RTS/CTS-based virtual carrier sensing isnecessary only when a hidden terminal is a critical problem.It has been shown that a hidden terminal rarely occurs inWiFi networks when the rate adaptation is enabled [29] andeven rarer under a light traffic load. Most WiFi devicesdisable the virtual carrier sensing by default, so as to savethe overhead caused by RTS/CTS. Hence, E-MiLi withoutRTS/CTS is more preferable in practice.

Association process. When E-MiLi coexists with legacyWiFi, the AP needs to discriminate them and prepend theM-preamble only for packets destined for E-MiLi-capableclients. The discrimination should be initialized during theassociation process, when a newly joining E-MiLi clientnotifies the AP about its capability, and subsequently theAP runs the address allocation algorithm to assign an

1452 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012

Fig. 17. FTP performance when data rate varies. “E-MiLi(short)” denotesE-MiLi with a short switching time (9:5 !s).

2. This issue was brought up by Sunghyun Choi during the ACMMobiCom 2011, Las Vegas, Nevada.

Page 13: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

address to it (and possibly reassign addresses to existingE-MiLi clients using the address allocation algorithm).

9 RELATED WORK

Energy-efficient protocols for WiFi. Energy efficiency haslong been a paramount concern for portable WiFi devices.Many MAC-level scheduling protocols have been proposedto reduce the energy wasted by IL. For example, NAPman[6] carefully isolates PSM clients’ traffic using an energy-aware fair scheduler, so as to reduce unnecessary IL causedby background traffic. SleepWell [30] further isolates thetraffic from different WLAN cells, by scheduling their wake-up time in a distributed TDMA manner. !PM [4] adopts amore fine-grained scheduler that aggressively puts clients tosleep even in between short packet intervals. E-MiLi can beintegrated with these and other MAC-level energy-savingsolutions, by adding the downclocked IL mode into theirstate machine (Section 6.2). E-MiLi can also work in CAM,thus overcoming the excessive delay typically seen in PSM-style protocols.

An alternative way of reducing the cost of IL is to wakeup the receiver on demand. The wake-on-wireless scheme[31] augments a secondary low-power radio for packetdetection, and triggers the primary receiver only when anew packet arrives. E-MiLi also adopts the philosophy ofon-demand packet processing. Its energy saving may be lessthan wake-on-wireless, because it needs to keep the analogcircuit active in IL. Its advantage is that no extra radio isrequired. In fact, it only requires a change of firmware tosupport the construction and detection of M-preamble, andadjustment of clock rate. E-MiLi can also be used withwake-on-wireless to optimize the power consumption of thesecondary radio.

Low-power listening (LPL) in sensor networks. Insensor networks, a popular MAC-layer energy savingmechanism is LPL, which is used by S-MAC [32], B-MAC[33] and many derivatives. Since sensor networks typicallyrun low-rate, small duty-cycle applications, LPL shifts morepower consumption to the transmitter side, thus reducingthe time spent in idle listening. Specifically, a receiverperiodically wakes up to detect packets from the transmitter,and the transmitter uses a long preamble that spans thatperiod to ensure detectability. Similar to the WiFi’s PSM,LPL is a sleep scheduling mechanism that reduces the ILtime, and can be enhanced by integrating with E-MiLi. Forexample, since E-MiLi reduces IL power, it can shorten thereceiver’s wake-up period, thereby shortening the transmit-ter’s preamble length and lowering its power consumption.

Packet detection. The general idea of correlation-basedpacket detection is not new. As mentioned in Section 5.3.2,the 802.11 OFDM PHY incorporates a preamble thatallows self-correlation-based detection. Its variants havealso been used in other software-radio implementations[34]. In E-MiLi, we have designed a new preamblemechanism that preserves the self-correlation propertyeven when it is downsampled. Cross-correlation-basedpacket detection (i.e., correlating the incoming signal witha known sequence) is an alternative way of detectingpackets [35], [36], but cannot detect downsampled signalsand is more susceptible to the frequency offset.

Dynamic voltage-frequency scaling (DVFS). DVFS is amature technology used in microprocessor design [7]. Itexploits the variance in processor load, lowering the voltageand clock rate when few tasks are pending, and raising itwhen the processor is heavily loaded. It has also beenproposed for Gigabit wireline links [37], and for audiosignal processing [8]. The key idea is to observe the peakfrequency of the incoming workload, and then limit theprocessor’s clock rate to that level.

DVFS has not been used for improving the energyefficiency for wireless radios, due mainly to a well-knownparadox: the radio should be activated only after detecting apacket, but to detect the packet, the radio must always beactive at its full sampling rate. We overcome this paradoxby separating packet detection and decoding, and perform-ing both at different rates. Our approach is partly inspiredby the experiments by Chandra et al. [3], who found WiFiNIC’s power consumption to scale linearly with thesampling bandwidth, and proposed the SampleWidthalgorithm to adjust the bandwidth according to the trafficload. SampleWidth uses the same clock rate for detectionand decoding, and can only adjust clock rate at a coarse-grained level, because the transmitter and the receiver mustagree on the same clock rate before packet transmissions.

10 CONCLUSION

We have presented E-MiLi, a novel mechanism forreducing the energy cost of IL that dominates the energyconsumption in WiFi networks. Our goal was to exercisefine-grained IL power control by adjusting clock ratewithout compromising packet-detection capability. Wemet this goal by devising a sampling-rate invariant packetdetector, which enables a downclocked radio to detectpackets with accuracy comparable to that of a full-clockedradio. We have also introduced an opportunistic down-clocking scheme to balance the overhead in changing clockrate and minimize its negative influence on networkperformance. Our experimental evaluation and trace-basedsimulation confirm the feasibility and effectiveness ofE-MiLi in real WiFi networks with different traffic patterns.

E-MiLi has wider implications for wireless design thanwhat we have explored in this paper. Its simple MAC/PHYinterface facilitates its integration with other carrier sensing-based wireless networks, such as ZigBee sensor networks.In addition, we only explored the benefits of downclockingin E-MiLi due to hardware limitation. By changing thevoltage along with clock rate, additional energy savings canbe achieved. This is a matter of our future inquiry.

ACKNOWLEDGMENTS

The work reported in this paper was supported in part bythe US Defense Advanced Research Projects Agency(DARPA) under Grant HR0011-10-1-0081. The main ideasof E-MiLi are part of a pending US patent.

REFERENCES

[1] G. Anastasi, M. Conti, E. Gregori, and A. Passarella, “802.11Power-Saving Mode for Mobile Computing in Wi-Fi Hotspots:Limitations, Enhancements and Open Issues,” Wireless Networks,vol. 14, no. 6, pp. 745-768, 2008.

ZHANG AND SHIN: E-MILI: ENERGY-MINIMIZING IDLE LISTENING IN WIRELESS NETWORKS 1453

Page 14: IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9 ...hy439/papers/emili-tmc-sep2012.pdf · tend to spend a large fraction of time in IL, due to MAC-level contention and network-level

[2] Y. Agarwal, R. Chandra, A. Wolman, P. Bahl, K. Chin, and R.Gupta, “Wireless Wakeups Revisited: Energy Management forVoIP over Wi-Fi Smartphones,” Proc. ACM MobiSys, 2007.

[3] R. Chandra, R. Mahajan, T. Moscibroda, R. Raghavendra, and P.Bahl, “A Case for Adapting Channel Width in Wireless Net-works,” Proc. ACM SIGCOMM, 2008.

[4] J. Liu and L. Zhong, “Micro Power Management of Active 802.11Interfaces,” Proc. ACM MobiSys, 2008.

[5] IEEE Std. 802.11, Wireless LAN Medium Access Control and PhysicalLayer Specifications, IEEE, 2007.

[6] E. Rozner, V. Navda, R. Ramjee, and S. Rayanchu, “NAPman:Network-Assisted Power Management for WiFi Devices,” Proc.ACM MobiSys, 2010.

[7] K. Flautner, S. Reinhardt, and T. Mudge, “Automatic PerformanceSetting for Dynamic Voltage Scaling,” Proc. ACM MobiCom, 2001.

[8] W.R. Dieter, S. Datta, and W.K. Kai, “Power Reduction by VaryingSampling Rate,” Proc. ACM/IEEE Int’l Symp. Low Power Electronicsand Design (ISLPED ’05), 2005.

[9] Ettus Research, “Universal Software Radio Peripheral (USRP),”http://www.ettus.com, 2012.

[10] A. Schulman, D. Levin, and N. Spring, “CRAWDAD Data Setumd/sigcomm2008,” 2008.

[11] C. Phillips and S. Singh, “CRAWDAD Data Set pdx/vwave,”2007.

[12] Atheros Comm., “AR5213 Preliminary Datasheet,” 2004.[13] Atheros Comm., “Power Consumption and Energy Efficiency of

WLAN Products,” 2004.[14] TI, Inc., “CC2420 Preliminary Datasheet,” 2004.[15] M. Zargari et al., “A Dual-Band CMOS MIMO Radio SoC for IEEE

802.11n Wireless LAN,” IEEE J. Solid-State Circuits, vol. 43, no. 12,pp. 2882-2895, Dec. 2008.

[16] B. McFarland, A. Shor, and A. Tabatabaei, “A 2.4 & 5 GHz DualBand 802.11 WLAN Supporting Data Rates to 108 Mb/s,” Proc.24th Ann. Technical Digest Gallium Arsenide Integrated Circuit (GaAsIC) Symp., 2002.

[17] J.T. Bevan et al., “An Integrated 802.11a Baseband and MACProcessor,” Proc. IEEE Int’l Solid-State Circuits Conf. (ISSCC) Digestof Technical Papers, 2002.

[18] P. Dutta, Y.-S. Kuo, A. Ledeczi, T. Schmid, and P. Volgyesi,“Putting the Software Radio on a Low-Calorie Diet,” Proc. NinthACM SIGCOMM Workshop Hot Topics in Networks (HotNets ’10),2010.

[19] A. Tabatabaei, K. Onodera, M. Zargari, H. Samavati, and D. Su, “ADual Channel "! ADC with 40MHz Aggregate Signal Band-width,” Proc. IEEE Int’l Solid-State Circuits Conf. (ISSCC) Digest ofTechnical Papers, 2003.

[20] Analog Devices, “AD9522 Data Sheet,” 2008.[21] P. Fan and M. Darnell, Sequence Design for Communications

Application. Research Studies, 1996.[22] J.K. Cavers,Mobile Channel Characteristics. Kluwer Academic, 2000.[23] Maxim, “MAX2831/MAX2832 2.4 GHz to 2.5 GHz 802.11g/b RF

Transceivers,” 2010.[24] Atheros Comm., “AR9280 Datasheet,” 2009.[25] M. Fujinami and T. Murakami, “PSM Extension for ns-2,” http://

nspme.sourceforge.net/index.html, 2012.[26] S. Gopal, S. Paul, and D. Raychaudhuri, “Investigation of the TCP

Simultaneous-Send Problem in 802.11 Wireless Local Area Net-works,” Proc. IEEE Int’l Conf. Comm. (ICC), 2005.

[27] S. Sankaran, M. Zargari, L. Nathawad, H. Samavati, S. Mehta, A.Kheirkhahi, P. Chen, K. Gong, B. Vakili-Amini, J. Hwang, S.-W.Chen, M. Terrovitis, B. Kaczynski, S. Limotyrakis, M. Mack, H.Gan, M. Lee, R. Chang, H. Dogan, S. Abdollahi-Alibeik, B.Baytekin, K. Onodera, S. Mendis, A. Chang, Y. Rajavi, S.-M. Jen, D.Su, and B. Wooley, “Design and Implementation of a CMOS802.11n SoC,” IEEE Comm. Magazine, vol. 47, no. 4, pp. 134 -143,Apr. 2009.

[28] A. Khattab, J. Camp, C. Hunter, P. Murphy, A. Sabharwal, andE.W. Knightly, “WARP: A Flexible Platform for Clean-SlateWireless Medium Access Protocol Design,” SIGMOBILE MobileComputing and Comm. Rev., vol. 12, pp. 56-58, 2008.

[29] M.Z. Brodsky and R.T. Morris, “In Defense of Wireless CarrierSense,” Proc. ACM SIGCOMM, 2009.

[30] J. Manweiler and R.R. Choudhury, “Avoiding the Rush Hours:WiFi Energy Management via Traffic Isolation,” Proc. ACMMobiSys, 2011.

[31] E. Shih, P. Bahl, and M.J. Sinclair, “Wake on Wireless: An EventDriven Energy Saving Strategy for Battery Operated Devices,”Proc. ACM MobiCom, 2002.

[32] W. Ye, J. Heidemann, and D. Estrin, “An Energy-Efficient MACProtocol for Wireless Sensor Networks,” Proc. IEEE INFOCOM,2002.

[33] J. Polastre, J. Hill, and D. Culler, “Versatile Low Power MediaAccess for Wireless Sensor Networks,” Proc. ACM Second Int’lConf. Embedded Networked Sensor Systems (SenSys), 2004.

[34] K. Jamieson and H. Balakrishnan, “PPR: Partial Packet Recoveryfor Wireless Networks,” Proc. ACM SIGCOMM, 2007.

[35] S. Sen, R.R. Choudhury, and S. Nelakuditi, “CSMA/CN: CarrierSense Multiple Access with Collision Notification,” Proc. ACMMobiCom, 2010.

[36] S. Sen, R.R. Choudhury, and B. Radunovic, “PHY-Assisted EnergyManagement for Mobile Devices,” Proc. ACM MobiSys PosterSession, 2010.

[37] L. Shang, L.-S. Peh, and N.K. Jha, “Dynamic Voltage Scaling withLinks for Power Optimization of Interconnection Networks,” Proc.IEEE Int’l Symp. High Performance Computer Architecture (HPCA),2003.

Xinyu Zhang received the BE degree in 2005from the Harbin Institute of Technology, China,and the MS degree in 2007 from the Universityof Toronto, Canada. He is currently workingtoward the PhD degree in the Department ofElectrical Engineering and Computer Science,University of Michigan. His research interestsare in the MAC/PHY codesign of wirelessnetworks, with applications in WLANs, WPANs,mesh networks, and white-space networks. He

was a recipient of the Best Paper Award at ACM MobiCom 2011. He is astudent member of the IEEE.

Kang G. Shin is the Kevin & Nancy O’Connorprofessor of computer science in the Departmentof Electrical Engineering and Computer Science,The University of Michigan, Ann Arbor. Hiscurrent research focuses on computing systemsand networks as well as on embedded real-timeand cyber-physical systems, all with an empha-sis on timeliness, security, and dependability. Hehas supervised the completion of 70 PhDdegrees and authored/coauthored more than

770 technical articles (more than 270 of these are published in archivaljournals), one textbook, and more than 20 patents or inventiondisclosures. He has also received numerous awards, including BestPaper Awards from the 2011 ACM International Conference on MobileComputing and Networking (MobiCom 2011), the 2011 IEEE Interna-tional Conference on Autonomic Computing, the 2010 and 2000USENIX Annual Technical Conferences, the 2003 IEEE Communica-tions Society William R. Bennett Prize Paper Award, and the 1987Outstanding IEEE Transactions of Automatic Control Paper Award. Hereceived the Research Excellence Award in 1989, the OutstandingAchievement Award in 1999, the Distinguished Faculty AchievementAward in 2001, and the Stephen Attwood Award in 2004 from TheUniversity of Michigan (the highest honor bestowed to MichiganEngineering faculty), a Distinguished Alumni Award of the College ofEngineering, Seoul National University, in 2002, the 2003 IEEE RTCTechnical Achievement Award, and the 2006 Ho-Am Prize in Engineer-ing (the highest honor bestowed to Korean-origin engineers). He haschaired several major conferences, including ACM MobiCom 2009,IEEE SECON 2008, ACM/USENIX MobiSys 2005, IEEE RTAS 2000,and IEEE RTSS 1987. He has served on several editorial boards,including the IEEE Transations on Parallel and Distributed Systems andACM Transactions on Embedded Systems. He has also served or isserving on numerous government committees, such as the US NationalScience Foundation Cyber-Physical Systems Executive Committee andthe Korean Government R&D Strategy Advisory Committee. He has alsocofounded a couple of startups. He is a fellow of the IEEE and ACM.

1454 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 11, NO. 9, SEPTEMBER 2012